Research On Improving Throughput Spatial Multitasking GPU

With the widespread use of GPUs,the issue of improving GPU resource utilization has also received widespread attention.Single application exclusive GPU or time-shared GPU can no longer make full use of a lot of resources on GPU.Space sharing GPU is a commonly used method to improve the utilization of GPU resources.Space-shared GPUs can take advantage of complementary resources between programs to improve GPU resource utilization.Existing air separation technologies cannot support dynamically changing the amount of program resources,and the monotonous resource allocation mode limits the application's ability to make full use of resources.This paper implements a process pool mechanism that can dynamically change the amount of program resources,making the program more flexible in scheduling resources.Based on this,this paper proposes two scheduling algorithms for two application scenarios to improve GPU resource utilization.This paper proposes a static scheduling algorithm for the scenario where multiple simple programs share the GPU.A simple program is a program that has only one type of task and a single demand for resources.The static scheduling algorithm reduces the bottleneck caused by insufficient resources by balancing various resources,improves the resource utilization of the GPU,and improves the overall GPU throughput by up to 27.9%.This paper proposes a dynamic scheduling algorithm for the scenario where multiple complex programs share the GPU.The demands on resources for complex programs are constantly changing.The dynamic scheduling algorithm improves the probability of complementary use of resources by increasing the probability of simultaneous execution of different types of tasks,improves the resource utilization of the GPU,and improves the overall throughput of the GPU by up to 13.6%.